AmbiDrop lets AI enhance speech from any mic array, no retraining needed
AmbiDrop uses Ambisonics and dropout to generalize speech enhancement to unseen arrays, boosting SI-SDR by 2-3 dB.
Most multichannel speech enhancement models are locked to specific microphone array geometries, failing when the hardware changes. Michael Tatarjitzky and Boaz Rafaely propose AmbiDrop, which sidesteps this limitation by converting any array’s recordings into the spherical harmonics domain using Ambisonics Signal Matching (ASM). A deep neural network is then trained on simulated Ambisonics data, with channel dropout applied to make the model robust to array-dependent encoding errors. This eliminates the need for massive multi-geometry datasets.
In experiments, AmbiDrop matched baseline performance on known arrays but significantly outperformed it on unseen layouts—achieving consistent gains in SI-SDR, PESQ, and STOI. The method demonstrates strong generalization, meaning a single model can handle arrays it has never encountered. Accepted at ICASSP 2026, AmbiDrop offers a practical path to deployment-agnostic speech enhancement, reducing hardware constraints for real-world audio systems.
- AmbiDrop uses Ambisonics Signal Matching to encode any microphone array into spherical harmonics, removing dependency on fixed geometry.
- Channel dropout during training ensures the model tolerates encoding errors from unseen array configurations.
- Achieves consistent improvements in SI-SDR, PESQ, and STOI over baselines on novel arrays, proving strong generalization without retraining.
Why It Matters
Enables speech enhancement that works on any microphone array, slashing hardware constraints and deployment costs for real-world audio systems.